Research map / Reinforcement learning / Visual and representation RL
Pre-Trained Image Encoder for Generalizable Visual Reinforcement Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use frozen ImageNet-pretrained ResNet features rather than a learned visual encoder | DrQ-v2 visual RL agent · DMC-GB video hard | episode return: +55% improvement on average | Helped |
| use frozen ImageNet-pretrained visual representations | visual RL agent · Drawer World | success rate: +24% boost on average | Helped |
| use Layer 2 features instead of Layer 4 features (ablation) | PIE-G · DMC-GB video hard, Walker Walk | episode return: 306 ± 31 -> 884 ± 20 | Helped |
| use Layer 2 features instead of Layer 1 features (ablation) | PIE-G · DMC-GB video hard, Walker Walk | episode return: 840 ± 32 -> 884 ± 20 | Helped |
| use Layer 2 features instead of Layer 3 features (ablation) | PIE-G · DMC-GB video hard, Walker Walk | episode return: 845 ± 27 -> 884 ± 20 | Helped |
| keep BatchNorm statistics updating during evaluation rather than using training statistics (ablation) | PIE-G · challenging generalization settings | generalization performance | Helped |
| replace ImageNet pre-training with CLIP pre-training (ablation) | PIE-G · DMC-GB video hard, Walker Walk | episode return: 600 ± 28 -> 615 ± 30 | No clear effect |
| replace ImageNet pre-training with Ego4D pre-training (ablation) | PIE-G · DMC-GB video hard, Walker Walk | episode return: 600 ± 28 -> 441 ± 15 | Hurt |
| finetune the pre-trained encoder instead of keeping it frozen (ablation) | PIE-G · DMC-GB video hard, Walker Walk | episode return: 600 ± 28 -> 455 ± 67 | Hurt |
| replace ImageNet pre-training with MoCo-v2 pre-training (ablation) | PIE-G · DMC-GB video hard, Walker Walk | episode return: 600 ± 28 -> 585 ± 30 | No clear effect |
| replace ResNet18 with ResNet34 or ResNet50 (ablation) | PIE-G · DMC-GB video hard | generalization performance | No clear effect |
Closest papers by meaning
- The Surprising Ineffectiveness of Pre-Trained Visual Representations for Model-Based Reinforcement Learning (NeurIPS 2024 · similarity 0.92)
- Focus On What Matters: Separated Models For Visual-Based RL Generalization (NeurIPS 2024 · similarity 0.92)
- Efficient Reinforcement Learning Through Adaptively Pretrained Visual Encoder (AAAI 2025 · similarity 0.90)
- VIGOR: Zero-Shot Visual Generalization via Latent-Space Consistency in Model-Based Reinforcement Learning (NeurIPS 2026 · similarity 0.89)
- Investigating Pre-Training Objectives for Generalization in Vision-Based Reinforcement Learning (ICML 2024 · similarity 0.88)
- A Simple Framework for Generalization in Visual RL under Dynamic Scene Perturbations (NeurIPS 2024 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.